I'm running a Kalman LLP5 algorithm within the MLTK to predict application crashes and account for trend and seasonality.
My search is:
| timechart span=1d sum(VOLUME) | predict "sum(VOLUME)" as prediction algorithm="LLP5" future_timespan="365" holdback="0" period=365 lower"95"=lower"95" upper"95"=upper"95" | `forecastviz(365, 0, "sum(VOLUME)", 95)`
However, as you can see, the algorithm predicts clear outliers just because they happened at the same time last year.
Is there a way to filter this out? And additionally, once it is filtered out, can I set an alert to tell me if the trend in the future does not follow this predicted trend or if there are outliers that fall well outside this confidence interval?
Additionally, if anyone knows of a better/more accurate way to run this forecast, I'd appreciate any suggestions.